############################################################################## # NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size # ############################################################################## # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.08 # ############################################################################## # This file is used to re-orangize all checkpoints (created by main-sss.py) # # into a single benchmark file. Besides, for each trial, we will merge the # # information of all its trials into a single file. # # # # Usage: # # python exps/NATS-Bench/sss-collect.py # ############################################################################## import os, re, sys, time, shutil, argparse, collections import torch from tqdm import tqdm from pathlib import Path from collections import defaultdict, OrderedDict from typing import Dict, Any, Text, List lib_dir = (Path(__file__).parent / '..' / '..' / 'lib').resolve() if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir)) from log_utils import AverageMeter, time_string, convert_secs2time from config_utils import dict2config from models import CellStructure, get_cell_based_tiny_net from nats_bench import pickle_save, pickle_load, ArchResults, ResultsCount from procedures import bench_pure_evaluate as pure_evaluate, get_nas_bench_loaders from utils import get_md5_file NATS_SSS_BASE_NAME = 'NATS-sss-v1_0' # 2020.08.28 def account_one_arch(arch_index: int, arch_str: Text, checkpoints: List[Text], datasets: List[Text]) -> ArchResults: information = ArchResults(arch_index, arch_str) for checkpoint_path in checkpoints: try: checkpoint = torch.load(checkpoint_path, map_location='cpu') except: raise ValueError('This checkpoint failed to be loaded : {:}'.format(checkpoint_path)) used_seed = checkpoint_path.name.split('-')[-1].split('.')[0] ok_dataset = 0 for dataset in datasets: if dataset not in checkpoint: print('Can not find {:} in arch-{:} from {:}'.format(dataset, arch_index, checkpoint_path)) continue else: ok_dataset += 1 results = checkpoint[dataset] assert results['finish-train'], 'This {:} arch seed={:} does not finish train on {:} ::: {:}'.format(arch_index, used_seed, dataset, checkpoint_path) arch_config = {'name': 'infer.shape.tiny', 'channels': arch_str, 'arch_str': arch_str, 'genotype': results['arch_config']['genotype'], 'class_num': results['arch_config']['num_classes']} xresult = ResultsCount(dataset, results['net_state_dict'], results['train_acc1es'], results['train_losses'], results['param'], results['flop'], arch_config, used_seed, results['total_epoch'], None) xresult.update_train_info(results['train_acc1es'], results['train_acc5es'], results['train_losses'], results['train_times']) xresult.update_eval(results['valid_acc1es'], results['valid_losses'], results['valid_times']) information.update(dataset, int(used_seed), xresult) if ok_dataset < len(datasets): raise ValueError('{:} does find enought data : {:} vs {:}'.format(checkpoint_path, ok_dataset, len(datasets))) return information def correct_time_related_info(hp2info: Dict[Text, ArchResults]): # calibrate the latency based on the number of epochs = 01, since they are trained on the same machine. x1 = hp2info['01'].get_metrics('cifar10-valid', 'x-valid')['all_time'] / 98 x2 = hp2info['01'].get_metrics('cifar10-valid', 'ori-test')['all_time'] / 40 cifar010_latency = (x1 + x2) / 2 for hp, arch_info in hp2info.items(): arch_info.reset_latency('cifar10-valid', None, cifar010_latency) arch_info.reset_latency('cifar10', None, cifar010_latency) # hp2info['01'].get_latency('cifar10') x1 = hp2info['01'].get_metrics('cifar100', 'ori-test')['all_time'] / 40 x2 = hp2info['01'].get_metrics('cifar100', 'x-test')['all_time'] / 20 x3 = hp2info['01'].get_metrics('cifar100', 'x-valid')['all_time'] / 20 cifar100_latency = (x1 + x2 + x3) / 3 for hp, arch_info in hp2info.items(): arch_info.reset_latency('cifar100', None, cifar100_latency) x1 = hp2info['01'].get_metrics('ImageNet16-120', 'ori-test')['all_time'] / 24 x2 = hp2info['01'].get_metrics('ImageNet16-120', 'x-test')['all_time'] / 12 x3 = hp2info['01'].get_metrics('ImageNet16-120', 'x-valid')['all_time'] / 12 image_latency = (x1 + x2 + x3) / 3 for hp, arch_info in hp2info.items(): arch_info.reset_latency('ImageNet16-120', None, image_latency) # CIFAR10 VALID train_per_epoch_time = list(hp2info['01'].query('cifar10-valid', 777).train_times.values()) train_per_epoch_time = sum(train_per_epoch_time) / len(train_per_epoch_time) eval_ori_test_time, eval_x_valid_time = [], [] for key, value in hp2info['01'].query('cifar10-valid', 777).eval_times.items(): if key.startswith('ori-test@'): eval_ori_test_time.append(value) elif key.startswith('x-valid@'): eval_x_valid_time.append(value) else: raise ValueError('-- {:} --'.format(key)) eval_ori_test_time = sum(eval_ori_test_time) / len(eval_ori_test_time) eval_x_valid_time = sum(eval_x_valid_time) / len(eval_x_valid_time) for hp, arch_info in hp2info.items(): arch_info.reset_pseudo_train_times('cifar10-valid', None, train_per_epoch_time) arch_info.reset_pseudo_eval_times('cifar10-valid', None, 'x-valid', eval_x_valid_time) arch_info.reset_pseudo_eval_times('cifar10-valid', None, 'ori-test', eval_ori_test_time) # CIFAR10 train_per_epoch_time = list(hp2info['01'].query('cifar10', 777).train_times.values()) train_per_epoch_time = sum(train_per_epoch_time) / len(train_per_epoch_time) eval_ori_test_time = [] for key, value in hp2info['01'].query('cifar10', 777).eval_times.items(): if key.startswith('ori-test@'): eval_ori_test_time.append(value) else: raise ValueError('-- {:} --'.format(key)) eval_ori_test_time = sum(eval_ori_test_time) / len(eval_ori_test_time) for hp, arch_info in hp2info.items(): arch_info.reset_pseudo_train_times('cifar10', None, train_per_epoch_time) arch_info.reset_pseudo_eval_times('cifar10', None, 'ori-test', eval_ori_test_time) # CIFAR100 train_per_epoch_time = list(hp2info['01'].query('cifar100', 777).train_times.values()) train_per_epoch_time = sum(train_per_epoch_time) / len(train_per_epoch_time) eval_ori_test_time, eval_x_valid_time, eval_x_test_time = [], [], [] for key, value in hp2info['01'].query('cifar100', 777).eval_times.items(): if key.startswith('ori-test@'): eval_ori_test_time.append(value) elif key.startswith('x-valid@'): eval_x_valid_time.append(value) elif key.startswith('x-test@'): eval_x_test_time.append(value) else: raise ValueError('-- {:} --'.format(key)) eval_ori_test_time = sum(eval_ori_test_time) / len(eval_ori_test_time) eval_x_valid_time = sum(eval_x_valid_time) / len(eval_x_valid_time) eval_x_test_time = sum(eval_x_test_time) / len(eval_x_test_time) for hp, arch_info in hp2info.items(): arch_info.reset_pseudo_train_times('cifar100', None, train_per_epoch_time) arch_info.reset_pseudo_eval_times('cifar100', None, 'x-valid', eval_x_valid_time) arch_info.reset_pseudo_eval_times('cifar100', None, 'x-test', eval_x_test_time) arch_info.reset_pseudo_eval_times('cifar100', None, 'ori-test', eval_ori_test_time) # ImageNet16-120 train_per_epoch_time = list(hp2info['01'].query('ImageNet16-120', 777).train_times.values()) train_per_epoch_time = sum(train_per_epoch_time) / len(train_per_epoch_time) eval_ori_test_time, eval_x_valid_time, eval_x_test_time = [], [], [] for key, value in hp2info['01'].query('ImageNet16-120', 777).eval_times.items(): if key.startswith('ori-test@'): eval_ori_test_time.append(value) elif key.startswith('x-valid@'): eval_x_valid_time.append(value) elif key.startswith('x-test@'): eval_x_test_time.append(value) else: raise ValueError('-- {:} --'.format(key)) eval_ori_test_time = sum(eval_ori_test_time) / len(eval_ori_test_time) eval_x_valid_time = sum(eval_x_valid_time) / len(eval_x_valid_time) eval_x_test_time = sum(eval_x_test_time) / len(eval_x_test_time) for hp, arch_info in hp2info.items(): arch_info.reset_pseudo_train_times('ImageNet16-120', None, train_per_epoch_time) arch_info.reset_pseudo_eval_times('ImageNet16-120', None, 'x-valid', eval_x_valid_time) arch_info.reset_pseudo_eval_times('ImageNet16-120', None, 'x-test', eval_x_test_time) arch_info.reset_pseudo_eval_times('ImageNet16-120', None, 'ori-test', eval_ori_test_time) return hp2info def simplify(save_dir, save_name, nets, total): hps, seeds = ['01', '12', '90'], set() for hp in hps: sub_save_dir = save_dir / 'raw-data-{:}'.format(hp) ckps = sorted(list(sub_save_dir.glob('arch-*-seed-*.pth'))) seed2names = defaultdict(list) for ckp in ckps: parts = re.split('-|\.', ckp.name) seed2names[parts[3]].append(ckp.name) print('DIR : {:}'.format(sub_save_dir)) nums = [] for seed, xlist in seed2names.items(): seeds.add(seed) nums.append(len(xlist)) print(' [seed={:}] there are {:} checkpoints.'.format(seed, len(xlist))) assert len(nets) == total == max(nums), 'there are some missed files : {:} vs {:}'.format(max(nums), total) print('{:} start simplify the checkpoint.'.format(time_string())) datasets = ('cifar10-valid', 'cifar10', 'cifar100', 'ImageNet16-120') # Create the directory to save the processed data # full_save_dir contains all benchmark files with trained weights. # simplify_save_dir contains all benchmark files without trained weights. full_save_dir = save_dir / (save_name + '-FULL') simple_save_dir = save_dir / (save_name + '-SIMPLIFY') full_save_dir.mkdir(parents=True, exist_ok=True) simple_save_dir.mkdir(parents=True, exist_ok=True) # all data in memory arch2infos, evaluated_indexes = dict(), set() end_time, arch_time = time.time(), AverageMeter() for index in tqdm(range(total)): arch_str = nets[index] hp2info = OrderedDict() full_save_path = full_save_dir / '{:06d}.pickle'.format(index) simple_save_path = simple_save_dir / '{:06d}.pickle'.format(index) for hp in hps: sub_save_dir = save_dir / 'raw-data-{:}'.format(hp) ckps = [sub_save_dir / 'arch-{:06d}-seed-{:}.pth'.format(index, seed) for seed in seeds] ckps = [x for x in ckps if x.exists()] if len(ckps) == 0: raise ValueError('Invalid data : index={:}, hp={:}'.format(index, hp)) arch_info = account_one_arch(index, arch_str, ckps, datasets) hp2info[hp] = arch_info hp2info = correct_time_related_info(hp2info) evaluated_indexes.add(index) hp2info['01'].clear_params() # to save some spaces... to_save_data = OrderedDict({'01': hp2info['01'].state_dict(), '12': hp2info['12'].state_dict(), '90': hp2info['90'].state_dict()}) pickle_save(to_save_data, str(full_save_path)) for hp in hps: hp2info[hp].clear_params() to_save_data = OrderedDict({'01': hp2info['01'].state_dict(), '12': hp2info['12'].state_dict(), '90': hp2info['90'].state_dict()}) pickle_save(to_save_data, str(simple_save_path)) arch2infos[index] = to_save_data # measure elapsed time arch_time.update(time.time() - end_time) end_time = time.time() need_time = '{:}'.format(convert_secs2time(arch_time.avg * (total-index-1), True)) # print('{:} {:06d}/{:06d} : still need {:}'.format(time_string(), index, total, need_time)) print('{:} {:} done.'.format(time_string(), save_name)) final_infos = {'meta_archs' : nets, 'total_archs': total, 'arch2infos' : arch2infos, 'evaluated_indexes': evaluated_indexes} save_file_name = save_dir / '{:}.pickle'.format(save_name) pickle_save(final_infos, str(save_file_name)) # move the benchmark file to a new path hd5sum = get_md5_file(str(save_file_name) + '.pbz2') hd5_file_name = save_dir / '{:}-{:}.pickle.pbz2'.format(NATS_SSS_BASE_NAME, hd5sum) shutil.move(str(save_file_name) + '.pbz2', hd5_file_name) print('Save {:} / {:} architecture results into {:} -> {:}.'.format(len(evaluated_indexes), total, save_file_name, hd5_file_name)) # move the directory to a new path hd5_full_save_dir = save_dir / '{:}-{:}-full'.format(NATS_SSS_BASE_NAME, hd5sum) hd5_simple_save_dir = save_dir / '{:}-{:}-simple'.format(NATS_SSS_BASE_NAME, hd5sum) shutil.move(full_save_dir, hd5_full_save_dir) shutil.move(simple_save_dir, hd5_simple_save_dir) # save the meta information for simple and full final_infos['arch2infos'] = None final_infos['evaluated_indexes'] = set() pickle_save(final_infos, str(hd5_full_save_dir / 'meta.pickle')) pickle_save(final_infos, str(hd5_simple_save_dir / 'meta.pickle')) def traverse_net(candidates: List[int], N: int): nets = [''] for i in range(N): new_nets = [] for net in nets: for C in candidates: new_nets.append(str(C) if net == '' else "{:}:{:}".format(net,C)) nets = new_nets return nets if __name__ == '__main__': parser = argparse.ArgumentParser(description='NATS-Bench (size search space)', formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument('--base_save_dir', type=str, default='./output/NATS-Bench-size', help='The base-name of folder to save checkpoints and log.') parser.add_argument('--candidateC' , type=int, nargs='+', default=[8, 16, 24, 32, 40, 48, 56, 64], help='.') parser.add_argument('--num_layers' , type=int, default=5, help='The number of layers in a network.') parser.add_argument('--check_N' , type=int, default=32768, help='For safety.') parser.add_argument('--save_name' , type=str, default='process', help='The save directory.') args = parser.parse_args() nets = traverse_net(args.candidateC, args.num_layers) if len(nets) != args.check_N: raise ValueError('Pre-num-check failed : {:} vs {:}'.format(len(nets), args.check_N)) save_dir = Path(args.base_save_dir) simplify(save_dir, args.save_name, nets, args.check_N)